Ureteral Orifice Detection in Ureteroscopic Images Based on Large-Kernel Convolutional Neural Networks and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Urology

Background:

• Ureteral orifice detection is crucial in ureteroscopic procedures.
• Existing methods may lack sufficient information modeling capacity for complex endoscopic imaging.

Purpose of the Study:

• To enhance large-kernel convolutional neural networks for improved information modeling.
• To develop a real-time ureteral orifice detection framework for ureteroscopic imaging.

Main Methods:

• A ConvNeXt backbone with SE and SCConv modules was utilized for feature extraction.
• Enhanced spatial excitation attention mechanisms were applied for feature fusion.
• Comparative experiments assessed accuracy, computational overhead, and inference latency against baseline models.

Main Results:

• The proposed network achieved a mean Average Precision (mAP@50) of 0.890 on 491 ureteroscopic images, surpassing baseline performance.
• The model demonstrated real-time performance with 50 FPS (20 ms per frame).

Conclusions:

• An improved deep learning framework for real-time ureteral orifice detection in endoscopic scenarios was developed.
• The system offers a balance between high detection accuracy and real-time efficiency.
• This method shows potential as a training tool for urologists.